Towards Last-layer Retraining for Group Robustness with Fewer Annotations

Summary

The paper tackles the aspect of last-layer retraining for better group robustness. The authors show that last-layer retraining can greatly improve worst-group accuracy with little worst-group data. Motivated by this, selective last-layer finetuning (SELF) is proposed to improve group robustness.

Strengths

1. The paper focuses on the interesting aspect of improving robustness via last-layer retraining under spurious correlation. The problem is specific and practical. 2. The experiments are well done. Also, the comparisons are consistent through out the paper. 3. The findings and proposed methods are well presented. The finding of "last-layer retraining can substantially improve WGA even when the reweighting dataset has only a small proportion of worst group data" is interesting.

Weaknesses

1. Novelty: (1) Last-layer retraining/finetuning is a common techqniue used in imbalanced/long-tailed learning and various properties were found, see [1][2]. This put doubt on the novelty of the claimed finding: "holding out a subset of training data to retrain the last layer can substantially outperform ERM on the entire dataset with no additional data, annotations, or computation for training". I wonder what are the differences if we treat tail classes in [1][2] as the worst group. (2) The proposed method: Disagreement-based methods were proposed and they were only adopted to the setting with spurious correlation. This adaptation is fine, but the performance in Table 4 cannot convince me since SELF cannot beat DFR in 2 out of 4 tasks. 2. Writing: (1) Figure 1(a) misses legends. For Figure 1(b), please note that y-axis means relative increase to avoid confusion. (2) The contribution part is too long. [1] Decoupling Representation and Classifier for Long-tailed Recognition. ICLR 2020 [2] BBN: Bilateral-Branch Network with Cumulative Learning for Long-Tailed Visual Recognition. CVPR 2020

Questions

Is worst-group accuracy a valid metric for evaluating robustness? A common phenomenon in long-tailed learning is that increasing tail-class accuracy often leads to decrease in head class accuracy. I recommend also adding overall accuracy as a metric for comparison. Also, WGA cannot reflect the uniformness of the overall performance.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The limitations are included in the appendix.

Summary

This paper proposes to collect a dataset from misclassifications or disagreements to fine-tune a classifier for improving sub-group accuracy in presence of spurious correlations.

Strengths

The topic that this paper attempts to address is important. The paper is well-written as well. There are some interesting findings in the paper as well such as finetuning using a dataset collected from disagreements could potentially improve group-accuracy.

Weaknesses

The level of novelty presented in the paper is insufficient. My primary concern stems from the method's reliance on group labels during model selection, also known as the validation set. This requirement renders these techniques nearly impractical in real-world scenarios, as defining groups is a non-trivial task even in datasets with just a hundred samples, and this becomes much more challenging for current datasets comprising billions of data points. The Waterbirds dataset utilized in this study has been identified to contain certain bugs, which were subsequently addressed in the MaskTune paper [1]. Consequently, the current values provided may be misleading. It is worth noting that the MaskTune paper, contrary to the assumption, does not incorporate any group labels during training or model selection (validation). It would be valuable for the authors to present the worst group accuracies achieved by SELF without the utilization of any training or validation group labels. Additionally, I encourage the authors to compare these results with the methods outlined in Tables 1 and 2 of [1]. [1] Asgari, Saeid, et al. "Masktune: Mitigating spurious correlations by forcing to explore." Advances in Neural Information Processing Systems 35 (2022)

Questions

Please refer to the Weaknesses section.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Please refer to the Weaknesses section.

Summary

The paper provides a detailed analysis of the DFR procedure [1]. The authors specifically consider the case when the group annotations are limited or not available during training, and show that it is still possible to largely remover the DFR performance. Specifically, the authors retrain the last layer on data with class balancing, and on data where examples are selected according to some rule (e.g. misclassified examples). The authors also provide many detailed ablations and understanding experiments that support their findings.

Strengths

**S1.** The authors provide new interesting insights into the DFR procedure. For example, the results in Figure 1 on the number of minority data needed are quite surprising: even with a small increase in the proportion of minority data it is possible to significantly improve worst-group performance. Also the results in Table 3 on the effectiveness of last-layer retraining without group rebalancing are thought-provoking. Finally, the results in Figure 3, which show that better group balancing does not always lead to better WGA. **S2.** The authors propose new practical methods (SELF), which build on DFR, but relax the requirement for group-balanced data. The results are discussed in Section 5. The authors also perform detailed ablations on the design decisions for the method, such as what cost function $c$ to use for selecting a subset of the data for last layer retraining. **S3.** SELF can work even when the reweighting data doesn’t have class labels, and request the labels on a small subset. Overall, I believe this paper provides new insights that expand our understanding of DFR and more generally training models robust to spurious features. The proposed methodology is also promising.

Weaknesses

I believe there is one important missing experiment, and several technical details that should be explained or corrected. I explain these in detail below. **W1.** Experiment: performance vs total number of group labeled examples As every other group robustness method, SELF requires group-labeled data to perform hyper-parameter tuning. My understanding is that in all of the experiments, half of the validation data is used for hyper-parameter tuning. In other words, it is not true that the method can work without any group annotations (which is also true of JTT, CnC, and other methods). Moreover, SELF seems to involve more hyper-parameters than DFR (which only tunes regularization strength): length of last layer retraining, size of the reweighting dataset, learning rate, etc. Consequently, I believe it is important to evaluate performance as a function of the total number of the group-labeled datapoints used by the method. In case of DFR, these datapoints would be used for last-layer retraining, and in case of SELF, they will only be used for hyper-parameter tuning. I think it may be reasonable to create a group-balanced validation set of size $n$, and plot WGA vs the size $n$. For example of a similar experiment, see Fig. 5 in [2]. I believe, this experiment is quite important for evaluating the methodological contribution of the paper. **W2.** Civilcomments I blieve, the results presented in Table 1 for CivilComments mix two different versions of the dataset. Specifically, JTT and CnC use the [Wilds version](https://wilds.stanford.edu/datasets/#civilcomments) with 16 overlapping groups, while DFR, RWY-ES, Group-DRO and the methods described in the paper use a version with 4 groups, where all the spurious attributes are combined together. As a result, there is a big discrepancy in performance. I would recommend either (1) reporting the results for all methods on the WILDS version of the dataset, or (2) removing the JTT and CnC entries and clarifying the difference in the versions of the data. **W3.** Waterbirds On waterbirds, the validation data is group balanced up to class balancing. As a result, by group-balancing, the authors are able to achieve DFR performance with class-balanced last-layer retraining (CB in Table 1). However, in this case the method is virtually identical to DFR. I think it would be more meaningful to do the experiment using a base reweighting dataset that has the same group distribution as the training data. This is the case for all other datasets.

Questions

**Q1. Early stopping:** What is the early stopping criterion for the models in Section 5? **Q2. Free lunch:** For the free-lunch results in Section 4.2 do you need regularization or early-stopping, optimized on validation WGA? Or is it quite robust to hyper-parameters? **Q3. Hyper-parameters:** Could you please list all hyper-parameters that a practitioner needs to tune for the SELF method? Finally, the paper [2] appears to be quite relevant, as it also attempts to automatically construct a reweighting dataset for DFR. However, [2] is a concurrent work, as it was published only after the submission of this paper. So there are no issues with the novelty of the paper. **References** [1] [_Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations_](https://openreview.net/forum?id=Zb6c8A-Fghk); P. Kirichenko, P. Izmailov, A. G. Wilson; ICLR 2023 [2] [_Simple and Fast Group Robustness by Automatic Feature Reweighting_](https://arxiv.org/abs/2306.11074); S. Qiu, A. Potapczynski, P. Izmailov, A. G. Wilson; ICML 2023

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Limitations are adequately discussed by the authors.

Summary

This work tackles an important problem of preventing the reliance of neural networks on spurious correlations. It builds on top of the work [1] primarily by using last layer re-training on class balanced held out dataset without the need for group annotations. They also additionally propose a simple but effective method SELF wherein the samples in disagreement with regularized models in the prediction with class annotations can be used for re-training the last layer. This method is particularly useful when group imbalance is large. Results are shown on Waterbirds, CelebA, MultiNLI and CivilComments datasets and performance is comparable to DFR [1].

Strengths

(1) The paper is well written and does very good analysis with ablation studies. (2) Makes an important observation that it is ok if re-weighting dataset has small proportion of worst group data. This proves that the class balancing is more important than the group balancing. (3) The paper addresses some of the practical challenges (in terms of computation and training overhead) by eliminating need to annotated group labels which is expensive, need to train twice, need to train the original ERM on class-balanced dataset. (4)The provided solutions are simple, and will not introduce any implementation or annotation overhead for already trained large scale models. The paper addresses the limitation of accuracy gap when the group imbalance is high and proposes a solution for the same with SELF technique.

Weaknesses

(1)The observation of class-balancing importance and re-evaluation of group balancing’s role in spurious correlation is a valuable contribution. However, solution in itself lacks novelty as it is not proposing a new algorithm per se. It seems very similar to [1] "Last layer re-training is sufficient for robustness to spurious correlations" with lesser constraints in terms of re-training dataset. (2) This particular line under section 5 “In addition to the balance of the reweighting, dataset, it is likely that characteristics of the specific data selected also contribute to SELF results. “ is unclear. Can you provide some qualitative examples from Celeb A where drop-out SELF disagreement does better than misclassification. More insights on what aspects of data makes the dropout SELF better would be good. Similarly, more qualitative example on the “uncertain data points” selected by the SELF-disagreement would provide more valuable insights. A lot of the observations are not supported by reasoning. (3) Since the key claim is the advantage of not needing group labels annotations - ablation studies on what is the overhead caused for obtaining group annotations (probably annotation time) will help quantify the effectiveness of the proposed method. Similarly ablation studies on the training compute time / mem consumotion between [1] and your method can help quantify the gains if any in training ERM on 95% split and using a MLP instead of Logistic regression. References: [1] "Last layer re-training is sufficient for robustness to spurious correlations"

Questions

Covered in weaknesses.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

A lot of the questions posed by the paper are still not addressed or analyzed . For e.g., It would be great to see more intuitive and theoretical reasoning on fundamental observations like: why last layer re-training and retraining on held out small class balanced data is improving the group robustness. Though experimental evidence has been provided, further analysis on the theoretical reasoning behind this would make a significant contribution in this subject.

Update

Thank the authors for their detailed response. Most of my concerns are resolved and I will raise my score.

Thank you for the rebuttal

_"Our main goal in this paper is not to beat the results in [1, 7], but rather to investigate the surprising performance of last-layer retraining in a more restrictive setting..."_ Given the stopping criteria for training (why stop at epochs 2, 4, 10?) and using labelled data for model selection, I don't see much novelty compared to [29], hence I keep my score.

Thank you for the rebuttal

Dear authors, thank you for the rebuttal and clarifications. I am mostly satisfied with the response. For the experiment on the dependence on the number of labeled validation points, it would be great to see a comparison with DFR specifically, to make sure that the proposed method improves label efficiency compared to DFR. That being said, I believe this paper provides a nice contribution, and I vote for acceptance.